Codex 校核（2026-10-09） Forge 当前 README把 8B 的结果明确限定为 v0.7.0 的26场景评测：从个位数到84%；Sonnet4.6的85%→98%来自v0.6.0，因成本未在v0.7.0重测。下文53%→99%不能作当前通用成功率，且工具格式可解析不等于任务或安全正确。Forge也可通过 required_steps、prerequisites、terminal_tool约束工作流；“只解决格式问题”遗漏了其响应校验和执行约束，但它仍不等于完整安全边界或多代理编排器。Stars／增长率与“市场需求”判断保留七月时点，未经独立市场验证。
Forge — Agent Reliability Guardrails Layer
Core Info
- Stars: 2,154⭐ (+115.4% growth 🚨)
- HN: 687pts / 252comments
- URL: https://github.com/antoinezambelli/forge
What It Does
Reliability guardrails that take an 8B local model from 53%→99% on agentic tool-calling tasks. Drop-in proxy that fixes tool call formats, rescues parsing errors, validates responses.
Key Technical Features
- Proxy Mode: Drop-in OpenAI/Anthropic compatible proxy — zero code integration with any harness
- Three Operation Modes: proxy server (lowest friction), WorkflowRunner, guardrails middleware
- Performance: Sonnet 4.6 85%→98%, 8B model single digits→84%
- Covers: Tool format fixing, retry prompting, response validation
Strategic Significance
- Validates “agent reliability layer” as a formal product category (not transitional trend)
- +115% growth in ~2 weeks confirms market pull
- Only solves tool-calling format issues — does NOT cover security/privilege/memory
Relevance to Hermes
- Forge + Hermes = tool reliability + security/memory = complete agent infrastructure
- Proxy mode integration pattern worth studying for Hermes guardrails
- Hermes safe-agent skill pack covers security angle Forge misses